Why Partitioning Matters: Revealing Overestimated Performance in WiFi-CSI-Based Human Action Recognition

Abstract

Human action recognition (HAR) based on WiFi channel state information (CSI) has attracted growing attention due to its contactless, privacy-preserving, and cost-effective nature. Recent studies have reported promising results by leveraging deep learning and image-based representations of CSI. However, methodological flaws in experimental protocols, particularly improper dataset partitioning, can lead to data leakage and significantly overestimate model performance. In this paper, we critically analyze a recently published WiFi-CSI-based HAR approach that converts CSI measurements into images and applies deep learning for classification. We show that the original evaluation relied on random data splitting without subject separation, causing substantial data leakage and inflated results. To address this, we reimplemented the method using subject-independent partitioning, which provides a realistic assessment of generalization ability. Furthermore, we conduct a quantitative study of post-training quantization under both correct and flawed partitioning strategies, revealing that methodological errors can conceal the true performance degradation of compressed models. Our findings demonstrate that evaluation protocols strongly influence reported outcomes, not only for baseline models but also for engineering decisions regarding model optimization and deployment. Based on these insights, we provide guidelines for designing robust experimental protocols in WiFi-CSI-based HAR to ensure methodological integrity and reproducibility.

1. Introduction

Human action recognition (HAR) has gained significant attention across domains such as healthcare [1], smart homes [2], and human–computer interaction [3] due to its potential for contactless and privacy-preserving sensing. Among the various modalities, WiFi-based HAR stands out because it leverages existing wireless infrastructure and does not rely on wearable sensors or cameras. The fine-grained channel state information (CSI) extracted from WiFi signals captures subtle motion-induced variations in the wireless channel, enabling the recognition of complex human activities using commodity hardware.

Recent studies have demonstrated impressive performance by transforming CSI data into image-like representations and applying deep learning architectures. However, many of these works suffer from methodological flaws—most critically, improper dataset partitioning that mixes data from the same subjects across training and testing sets. This practice leads to data leakage, allowing the model to memorize subject-specific patterns rather than learning to generalize to unseen individuals. As a result, reported accuracies are often unrealistically high and fail to reflect performance in real-world deployments.

This paper reexamines a representative WiFi CSI-based HAR method—the paper by Shahverdi et al. [4]—and shows that such flawed partitioning can distort not only baseline accuracy but also downstream evaluations such as model compression through post-training quantization. By reimplementing the method with subject-independent partitioning, we reveal the true generalization capability of the model and quantify how leakage can conceal the degradation introduced by quantization. Our findings underline that evaluation protocols—especially dataset partitioning—play a decisive role in determining the credibility and reproducibility of CSI-based HAR research.

1.1. Contributions

The main contributions of this paper are summarized as follows:

• Critical analysis of an existing WiFi-CSI-based HAR method that applies image-based processing and deep learning. We show that the original data partitioning strategy used in that work introduces significant data leakage.
• Reimplementation of the original approach with a correct, subject-independent partitioning strategy and rigorous evaluation, demonstrating the actual generalization performance.
• Quantitative analysis of post-training quantization under both correct and incorrect data partitioning, showing that methodological flaws can mask substantial performance degradation—in our case, resulting in an overestimation of about 32% in F1-score when data from the same subjects appeared in both training and test sets.
• Comprehensive discussion and guidelines for designing robust evaluation protocols in WiFi-CSI-based HAR, helping future studies avoid similar methodological pitfalls.

While the vast majority of existing WiFi CSI-based HAR studies appear to employ non-subject-independent evaluation protocols, providing a precise quantitative estimate of this prevalence would require an exhaustive reimplementation and verification of numerous prior works, which lies beyond the scope of this study. Our intent here is to highlight the methodological risks and demonstrate their tangible impact through a representative example. We hope this work encourages the community to undertake a broader quantitative assessment of this issue in future research.

1.2. Structure of the Paper

The remainder of this paper is organized as follows. Section 2 reviews related work with a focus on previous studies in WiFi-based human activity recognition and issues of data leakage. Section 3 introduces the necessary preliminaries, including definitions. Section 4 describes the materials and methods, with particular emphasis on the dataset preparation and the experimental setup. Section 5 presents the experimental results and analysis, where we compare the outcomes obtained under different data partitioning strategies and examine the impact of additional factors such as quantization. Section 6 discusses the broader implications of our findings, highlighting methodological limitations and lessons for future research. Finally, Section 7 concludes the paper by summarizing the main contributions and outlining potential directions for further work.

2. Related Work

This section reviews research relevant to our study from two perspectives. First, we summarize the main approaches and recent advancements in WiFi-CSI-based human activity recognition. Second, we discuss studies on data leakage in machine learning research, emphasizing its impact on model evaluation and the importance of rigorous dataset partitioning strategies.

2.1. WiFi-CSI-Based HAR

WiFi CSI has increasingly been leveraged for HAR in recent years due to its non-invasive capabilities and the growing ubiquity of WiFi networks. The ability of CSI to encapsulate detailed spatial and temporal data arising from human movements makes it a promising tool for accurately discerning different human actions [5].

Pioneering WiFi-CSI-based HAR frameworks utilized statistical values, e.g., average, variance, or median, as features of absolute CSI time-series and trained classifiers to predict a human action based on a newly given CSI signal. For instance, Wang et al. [6] aggregated 30 CSI subcarrier vectors into one vector by applying weighted moving average. Since the authors implemented a WiFi-based human fall detection system, a one-class support vector machine (SVM) [7] was used to classify the aggregated feature vector into “fall” and “not-fall” categories. A similar approach was followed by Al-qaness [8], but six features were selected both from CSI amplitude and CSI phase, and random forest [9] was utilized as a classification module. In contrast, Wang et al. [10] analyzed the moving variance of the CSI amplitude to identify human actions. In [11], discrete wavelet transform was used for feature extraction from CSI, and similarly to Wang et al. [6], a SVM was applied for HAR based on the extracted features.

In recent years, deep learning methods have become the dominant approach for WiFi-CSI-based human activity recognition, offering automated feature extraction. Deep-learning-based approaches for WiFi-CSI-based HAR can generally be categorized into two groups. The first group directly maps raw or preprocessed CSI time-series to activity classes using models, such as RNNs, LSTMs, or 1D CNNs, while the second group transforms CSI sequences into time–frequency or spatial representations, effectively converting them into images and leveraging architectures originally developed for image classification. For instance, Yousefi et al. [12] applied short-time Fourier transform on a certain number of denoised CSI samples to extract features. On the top of these features, a long short-term memory (LSTM) [13] was trained to predict human actions. In [14], the authors considered an advanced version of LSTM, i.e., bi-directional LSTM. Schäfer et al. [15] combined multiple LSTMs with an SVM to carry out CSI-based HAR. In contrast, Shen et al. [16] trained a 1D-CNN on CSI data to implement a keystroke recognition system. Zhang and Jiao [17] proposed a novel approach in which denoised CSI signals are transformed into images using Gramian angular field [18] representation, and these images are subsequently classified with a 2D CNN. In addition, the authors compared the performance of several pretrained architectures with that of a custom architecture trained from scratch. Moshiri et al. [19] also employed a 2D CNN but, instead of using a Gramian transform, constructed a resized pseudo-image directly from the CSI signals using the MATLAB’s imagesc function. Similarly, Jawad et al. [20] generated MATLAB-based pseudo-images, but after first applying Hampel filtering to the CSI signals, and adapted pretrained networks (VGG19 [21], AlexNet [22], SqueezeNet [23]) for the classification task.

For readers seeking a broad overview of HAR across different sensing modalities—including wearable sensors, vision-based systems, radar, and WiFi-based approaches—we refer to comprehensive survey papers that cover the entire field [24,25,26,27]. In contrast, other references [28,29] focus on the categorization of WiFi-CSI-based HAR methods.

2.2. Data Leakage in Machine Learning Research

Data leakage in machine learning research refers to the inadvertent use of information in the training process that would not be available at the time of prediction, leading to overly optimistic and unreliable model performance. This phenomenon critically undermines the ability of models to generalize to unseen data, causes biased decision-making, and produces misleading analytical insights, significantly affecting the validity and applicability of machine learning results.

Data leakage occurs when a machine learning model gains access to external or future information during training that inflates its predictive accuracy artificially. This can happen due to technical errors such as applying data preprocessing steps (like scaling, normalization, or imputation) to the entire dataset before splitting into training and testing sets, thereby exposing data from the test set to the training process. Other common causes include merging external datasets that contain direct or indirect information about the target variable, improper feature selection from the full dataset rather than only the training portion, and repeated instances of the same subject or samples across training and test sets. Additionally, vulnerabilities in data integration pipelines—such as insecure data transmission, misconfigured access controls, or legacy system interactions—can introduce subtle leakage points, especially in enterprise or production environments where data flows through various systems before model training [30,31].

Kaufman et al. [32] addressed the problem of data leakage in data mining, where illegitimate information about the prediction target is unintentionally introduced into the modeling process. Further, the authors provided a formal definition of leakage, distinguishing between feature leakage (e.g., using future or target-derived features) and training sample leakage (e.g., dependencies between training and test sets). The authors presented methods to avoid leakage, such as tagging data for legitimacy and enforcing a strict learn–predict separation, and discussed detection approaches based on exploratory data analysis and unexpected model performance. They also showed how causal modeling concepts help interpret complex leakage scenarios and emphasized that fixing leakage after it occurs is often difficult. Kapoor and Narayanan [33] examined how data leakage contributes to widespread reproducibility issues across scientific fields that apply machine learning. Based on a review of 22 studies covering 17 disciplines, the authors found data leakage in at least 294 papers, often leading to overly optimistic results. They introduced a taxonomy of eight types of leakage (i.e., [L1] a lack of a clean separation of training and test datasets, [L1.1] no test set, [L1.2] pre-processing on training and test sets, [L1.3] feature selection on training and test sets, [L1.4] duplicates in datasets, [L2] model use features that are not legitimate, [L3] a test set that is not drawn from the distribution of scientific interest, [L3.1] temporal leakage, [L3.2] non-independence between training and test samples, and [L3.3] a sampling bias in the test distribution) and propose model info sheets as a practical tool for researchers to detect and prevent such issues. A case study on civil war prediction demonstrated that, once leakage is corrected, complex machine learning models perform no better than traditional statistical methods, highlighting the need for stricter methodological practices to ensure reliable scientific findings.

3. Preliminaries

This section provides the background knowledge necessary to understand the proposed method. First, we introduce WiFi channel state information, which represents the physical-layer characteristics of wireless communication channels and is widely used for contactless human activity recognition. Next, we present Canny edge detection, a classical image processing technique that plays an important role in enhancing feature extraction from CSI-derived representations.

3.1. WiFi Channel State Information

In WiFi-based HAR, the most commonly used physical-layer feature is the channel state information (CSI), which characterizes the wireless channel at the time of transmission. Modern WiFi standards based on orthogonal frequency division multiplexing (OFDM) estimate a complex channel frequency response for each subcarrier [34]. In multiple-input multiple-output (MIMO) systems, the CSI describes how the transmitted signal is affected by propagation phenomena such as reflection, scattering, and fading. Knowing the CSI enables adaptive transmission strategies to achieve higher data rates and reliable communication [35]. If $T(f_n,t)$ and $R(f_n,t)$ represent the transmitted and received signals at the nth subcarrier frequency $f_n$ and time t, respectively, the received signal can be expressed as

$$R(f_n,t)=T(f_n,t)\times CSI(f_n,t)+W,$$

(1)

where W denotes the channel noise. The CSI is represented as a complex matrix:

$${\mathbf{H}}_k=\left(\begin{array}{ccc}{h}_{1,1}& \cdots & h_{1,r}\\ ⋮& \ddots & ⋮\\ h_{t,1}& \cdots & h_{t,r}\end{array}\right)$$

(2)

where t and r are the numbers of transmit and receive antennas, respectively. Each element is a complex value:

$$h_{t,r}=\left|h_{t,r}\right|e^{j\angle h_{t,r}}$$

(3)

containing both the amplitude $|h_{t,r}|$ and the phase $\angle h_{t,r}$ of the propagation path between the tth transmit antenna and rth receive antenna.

Compared with coarse-grained metrics, such as the received signal strength indicator (RSSI), CSI provides fine-grained information by measuring the channel frequency response at each subcarrier and antenna pair. This enables CSI to capture subtle multi-path variations caused by human motion, such as arm swings, walking, or even chest movements during breathing. These small changes induce measurable perturbations in both the amplitude and phase of CSI, which can be exploited to distinguish different activities without the need for wearable sensors or cameras. Because CSI can be obtained using commodity WiFi hardware (e.g., Intel 5300 NICs or Raspberry Pi boards with Nexmon firmware), it offers a low-cost and privacy-preserving solution for contactless human activity recognition [36]. Each captured WiFi frame yields a set of CSI values across subcarriers, antennas, and time. These values are typically preprocessed (e.g., noise filtering, normalization, dimensionality reduction) and then fed into machine learning or deep learning models to classify human activities accurately.

3.2. Canny Edge Detection

Edge detection is a fundamental technique in computer vision, aiming to identify significant local changes in intensity within an image, which often correspond to object boundaries or salient structural information [37]. Among various edge detection algorithms, the Canny edge detector, proposed by John F. Canny in 1986, remains one of the most widely used due to its optimal performance under a set of criteria [38]. The Canny edge detection algorithm involves the following sequential steps.

• Noise reduction. The input image I is smoothed using a Gaussian filter $G_{\sigma }$ to suppress noise:

  $$I_s(x,y)=I(x,y)\ast G_{\sigma }(x,y),$$

  (4)

  where ∗ denotes the convolution and

  $$G_{\sigma }(x,y)={\displaystyle \frac{1}{2\pi {\sigma }^2}}{e}^{-{\displaystyle \frac{x^2+y^2}{2{\sigma }^2}}}$$

  (5)

  with $\sigma$ being the standard deviation of the Gaussian kernel.
• Gradient calculation. The gradients in the x and y directions—$G_x$ and $G_y$—are computed, typically using Sobel operators. The gradient magnitude M and direction $\Theta$ at each pixel are then calculated as

  $$M=\sqrt{G_x^2+G_y^2},$$

  (6)

  $$\Theta =arctan\left({\displaystyle \frac{G_y}{G_x}}\right).$$

  (7)
• Non-maximum suppression. The algorithm suppresses all gradient magnitudes that are not local maxima along the gradient direction, resulting in thin edges.
• Double thresholding.

  $$M(x,y)=\left\{\begin{array}{cc}\mathrm{strong},\hfill & M(x,y)\ge T_H\hfill \\ \mathrm{weak},\hfill & T_L\le M(x,y)<T_H\hfill \\ \mathrm{zero},\hfill & M(x,y)<T_L,\hfill \end{array}\right.$$

  (8)

  where $T_H$ and $T_L$ are the high and low thresholds, respectively.
• Edge tracking by hysteresis. Weak edges are retained only if they are connected to strong edges, ensuring edge continuity and discarding isolated responses.

3.3. Post-Training Quantization

Modern deep neural networks often require millions of parameters and substantial computational resources, making their deployment challenging on embedded systems and edge devices with limited memory and processing power. Quantization addresses this issue by representing model parameters and, in some cases, intermediate activations with reduced numerical precision [39].

In PTO, quantization is applied to a fully trained model without modifying the training process. Let $\mathbf{W}\in {\mathbb{R}}^n$ denote a vector of trained model weights. In PTQ, each weight $w\in \mathbf{W}$ is mapped to a discrete set of quantized values $\widehat{\mathbf{W}}$, typically represented by integers of lower bit-width, such as 8-bit signed integers:

$$\widehat{w}=round\left({\displaystyle \frac{w}{s}}\right),$$

(9)

$$s={\displaystyle \frac{max\left(W\right)-min\left(W\right)}{2^b-1}},$$

(10)

where b is the number of bits (e.g., $b=8$), and s is a scale factor used for dequantization during inference:

$$w\approx s·\widehat{w}.$$

(11)

This transformation reduces the model size approximately by a factor of ${\displaystyle \frac{32}{b}}$ compared to standard 32-bit floating-point representations.

Quantization can be applied to different parts of the network:

• Weights only (static quantization): All model parameters are quantized, while activations remain in floating point.
• Weights and activations (full quantization): Both parameters and intermediate values are quantized, which further reduces latency and memory requirements.

However, quantization introduces quantization error, defined as

$$E_q=\left|\right|\mathbf{W}-s·\widehat{\mathbf{W}}{\left|\right|}_2,$$

(12)

which can degrade the model’s prediction accuracy. The severity of this degradation depends on model structure, input data distribution, and quantization parameters.

PTQ is attractive because it does not require retraining and can be applied to any pre-trained network, but its effectiveness must be evaluated on realistic benchmarks. Inaccurate evaluation protocols (e.g., those with data leakage between training and test sets) can underestimate the actual performance loss caused by quantization, leading to overly optimistic conclusions about model deployability on constrained devices.

4. Materials and Methods

This section details the resources and methodologies used in this study. First, we introduce the dataset employed in our experiments, including its collection process and structure. Next, we describe the proposed method for processing raw CSI data and training the classification model. Finally, we present the detected data leakage issue in the original experimental setup, highlighting its implications for model evaluation and generalization.

4.1. Applied Database

The CSI-HAR dataset was collected by Moshiri et al. [19] for WiFi-based HAR using CSI. Data acquisition was performed in an indoor environment with a Raspberry Pi 4 running the Nexmon CSI tool [40] and a TP-Link Archer C20 access point operating at 5 GHz and 20 MHz bandwidth. Seven daily activities—walking, running, falling, lying down, sitting down, standing up, and bending— were carried out by three volunteers of different ages, each repeated 20 times, resulting in 420 activity samples. CSI data were captured in monitor mode as raw packets and segmented based on activity duration, yielding matrices with 52 subcarrier columns and 600–1100 time samples depending on the action performed. No pre-processing or filtering was applied to preserve signal characteristics; instead, the matrices were normalized. The dataset is publicly available on GitHub and is intended to support research on low-cost, real-time human activity recognition for healthcare and smart home applications. Figure 1 illustrates the actions of CSI-HAR dataset [19] with pseudocolor images.

4.2. Proposed Method

Figure 2 illustrates Shahverdi et al.’s [4] CSI data preprocessing steps. First, principal component analysis (PCA) is applied to the raw CSI data suppress noise by retaining only the most informative components. This step helps highlight dominant patterns in the CSI variations while discarding redundant or irrelevant information. The normalized, reduced-dimension data are then transformed into pseudocolor RGB images using MATLAB R2022b’s imagesc functione, after which the Canny edge detection algorithm is applied to emphasize structural patterns in the signal representation. Finally, the edge-detected images are resized to the required input dimensions $(64\times 64)$ for the deep learning model and stored with their corresponding activity labels.

Figure 3 depicts the structure of the CNN implemented by Shahverdi et al. [4]. It consists of two convolutional layers, each followed by leaky rectified linear unit (ReLU) activations to introduce non-linearity and mitigate vanishing gradient problems. Max pooling layers (kernel size of $2\times 2$ and stride of 1) succeed the convolutional stages to reduce spatial dimensions and introduce translation invariance, helping the model focus on feature presence rather than precise location. After convolution and pooling, the resulting feature maps are flattened into a one-dimensional vector and passed through two fully connected dense layers, which synthesizes the extracted features. The final output layer has neurons equal to the number of activity classes, delivering the probabilistic classification of observed human activities. Additional regularization techniques such as dropout (dropout rate was set to 0.2 for the convolutional layers and 0.1 for the fully connected layers) and batch normalization are implicitly integrated to prevent overfitting and ensure stable learning, particularly essential given the limited size and noisy nature of CSI datasets.

As the original publication [4] did not provide detailed specifications of the training procedure, the Adam optimizer was adopted in our implementation. The remaining training parameters are summarized in

Table 1. Parameter settings.

Parameter	Value
Loss function	Cross-entropy
Optimizer	Adam [43] (${\beta }_1=0.9$, ${\beta }_2=0.999$, $ϵ={10}^{-8}$)
Learning rate	0.001
Weight decay	0.0
Batch size	128
Number of epochs	40

. The implementation of Shahverdi et al.’s method [4] was carried out using Python 3.11.5 and PyTorch 2.5.1 [42]. The hardware configuration is provided separately in

Table 2. Computer configuration.

Parameter	Value
Computer model	STRIX Z270H Gaming
Operating system	Windows 10
CPU	Intel(R) Core(TM) i7-7700K CPU 4.20 GHz (8 cores)
Memory	15 GB
GPU	NVIDIA GeForce GTX 1080

.

4.3. Detected Data Leakage

The presence of data leakage was first suspected when we saw in the paper of Shahverdi et al. [4] that the model achieved an unusually high test accuracy, comparable to the performance typically observed on well-structured natural image datasets such as CIFAR-10 [44,45]. Given that CSI-derived images are inherently noisy, low-contrast, and contain subtle, overlapping motion patterns rather than clearly separable object categories, such performance levels appeared implausible. This discrepancy suggested that the model might have exploited subject-specific cues rather than learning generalizable activity features. Subsequent inspection of the dataset confirmed that samples from the same individuals appeared in both the training and test sets, validating the presence of data leakage.

A critical methodological issue was identified in the way the dataset was partitioned into training and testing subsets. The authors did not perform the split on a per-subject basis; instead, samples from the same individual appear in both the training and the test sets—illustrated in Figure 4. Since the raw CSI signals inherently contain person-specific characteristics—such as body shape, walking style, or other idiosyncratic movement patterns—the model is able to learn and exploit these features rather than focusing solely on the activity patterns it is intended to recognize. This results in a substantial overestimation of classification performance, as the model is indirectly “seeing” information during training that it should not have access to at inference time.

Such subject-independent partitioning (shown in Figure 5) is a well-established requirement in human activity recognition research, precisely to avoid this form of leakage. By not enforcing it, the reported accuracy reflects the model’s ability to re-identify participants rather than to generalize to unseen individuals. Consequently, the results presented in the paper are likely inflated and may not be reproducible in real-world scenarios where the system must classify activities of new users. This issue represents a form of feature leakage, where unintended identity cues are leaked from the training set into the test set, undermining the validity of the experimental evaluation.

5. Experimental Results and Analysis

This section presents the experimental evaluation of the proposed approach. We first describe the metrics used to assess classification performance, ensuring a comprehensive and interpretable comparison across different experimental settings. Next, we report and analyze the numerical results, focusing on the impact of data partitioning strategies and other relevant factors, such as quantization, on model accuracy and generalization.

5.1. Evaluation Metrics

To assess the performance of the classification models, standard evaluation metrics are employed. Let $TP$, $TN$, $FP$, and $FN$ denote the number of true positives, true negatives, false positives, and false negatives, respectively. The metrics are defined as follows.

• Accuracy is the ratio of correctly classified instances to the total number of instances:

  $$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}}.$$

  (13)
• Precision is the proportion of correctly predicted positive samples among all samples predicted as positive:

  $$Precision={\displaystyle \frac{TP}{TP+FP}}.$$

  (14)
• Recall is the proportion of correctly predicted positive samples among all actual positive samples:

  $$Recall={\displaystyle \frac{TP}{TP+FN}}.$$

  (15)
• F1-score is the harmonic mean of precision and recall:

  $$F1-score=2·{\displaystyle \frac{Precision·Recall}{Precision+Recall}}.$$

  (16)

• Since the dataset used in this study is class-balanced, macro-averaged values of the above metrics provide a fair and reliable measure of classification performance. Nevertheless, it is still important to report precision, recall, and F1-score in addition to accuracy, as these metrics capture complementary aspects of performance and highlight the trade-offs between different types of classification errors.

5.2. Numerical Results

In this section, we report the numerical results obtained under two different data partitioning strategies. First, we consider the results based on the subject-independent split, which represents the correct experimental protocol and provides a realistic estimate of model performance on unseen users. Second, we examine the results obtained with the subject-overlapping split, as originally used in the paper, which inadvertently introduces data leakage by allowing samples from the same individuals to appear in both training and test sets. By comparing these two evaluation settings, we are able to quantify the extent to which the reported performance is inflated due to the flawed partitioning. In addition, we also analyze further factors influencing the results, such as the effect of model quantization.

Table 3. Reported and re-implemented classification results under different data partitioning strategies.

	Accuracy	F1-Score
Reported in [4]	0.92	0.92
Retrained w/o.r.t. humans	0.929	0.928
Retrained w.r.t. humans	0.607	0.604

summarizes the accuracy and F1-score reported in the original publication, along with the results obtained from our re-implementation. When trained and evaluated without separating subjects, our results closely match those reported in the paper, confirming the correctness of the re-implementation. However, when a subject-independent data split is applied, both accuracy and F1-score drop significantly (to 60.7% and 60.4%, respectively), clearly demonstrating that the original evaluation protocol suffers from data leakage and does not reflect true model generalization to unseen users.

Figure 6 and Figure 7 present the confusion matrices for the activity classification task under two different data partitioning strategies. In Figure 6, where the dataset was randomly split without separating subjects, the classification performance is consistently high across all activities, with several classes achieving perfect accuracy. This apparent robustness, however, reflects the model’s exploitation of subject-specific characteristics rather than its ability to generalize to unseen individuals. In contrast, Figure 7, which corresponds to the subject-independent split, reveals a substantial drop in classification performance, particularly for activities such as bend, fall, and run. Misclassifications become more frequent, and confusion between similar activities is clearly visible. This stark performance difference between the two settings confirms that the original evaluation protocol was affected by data leakage. The subject-independent results provide a more realistic measure of the model’s generalization capability, demonstrating that the performance reported in the original paper was significantly overestimated.

Figure 8 and Figure 9 illustrate the impact of post-training quantization with and without taking human-based separation into account during training. In Figure 8, where the network was trained without considering separation by human subjects (incorrect approach), the model achieves high accuracy in full precision (FP32) and maintains stable performance down to 5-bit quantization (around 92.9%). However, the accuracy drops significantly at 4 bits and below, with extreme degradation at 2 bits (46.4%) and 1 bit (14.3%). In contrast, Figure 9 shows the results when the training was correctly performed with subject-based separation. Here, the FP32 baseline accuracy is considerably lower (60.7%), and accuracy decreases gradually as quantization bits are reduced. The degradation pattern is smoother compared to Figure 6, but the overall accuracy is consistently lower, reaching only 52.9% at 3 bits and collapsing at 2 and 1 bits (17.1% and 14.3%, respectively). Overall, the comparison demonstrates that ignoring human-based separation during training may lead to artificially inflated accuracy in both FP32 and higher-bit quantization, while the correct subject-based separation yields lower but more reliable and realistic performance across quantization levels.

Figure 10 and Figure 11 compare the training dynamics under two different data partitioning strategies: without considering subject identities (Figure 10) and with subject-aware partitioning (Figure 11). In the subject-unaware setting, the model quickly achieves very high test accuracy, closely following the training accuracy curve. However, this is accompanied by suspiciously low test loss values, suggesting that the model may be exploiting subject-specific information rather than genuinely learning the underlying actions. In contrast, when subject identities are taken into account, the gap between training and test performance becomes more pronounced. Although the overall test accuracy is lower, the results provide a more realistic estimate of the model’s generalization ability, reflecting the true difficulty of the task. This comparison highlights the importance of subject-aware data partitioning in avoiding data leakage and ensuring a fair evaluation of model performance.

6. Discussion

Our results demonstrate that improper data partitioning, specifically the lack of subject separation, leads to severe data leakage in WiFi-CSI-based human action recognition. This not only inflates baseline performance but also obscures the true effects of engineering decisions, such as post-training quantization. In contrast, subject-independent partitioning yields lower but realistic performance, offering a fair assessment of generalization to unseen users [46]. These findings emphasize the need for rigorous evaluation protocols in WiFi-CSI-based HAR. Researchers should consistently apply subject-independent splits, report training dynamics, and consider complementary analyses such as confusion matrices or ground truth versus predicted plots. Establishing such practices will help prevent misleading conclusions, improve reproducibility, and ensure that future research in this domain produces results that are both reliable and practically relevant.

Although our analysis focuses on WiFi CSI-based human action recognition, similar methodological vulnerabilities may arise in other radio-frequency sensing modalities, such as radar or ultra-wideband systems. These technologies also capture user-dependent signal patterns, so improper partitioning can similarly inflate performance estimates. Addressing such cross-modal reproducibility concerns would help ensure methodological integrity across the broader RF sensing community.

7. Conclusions

This paper critically examined a WiFi-CSI-based human action recognition method and revealed that subject-unaware data partitioning introduces severe data leakage, leading to unrealistic performance estimates. By reimplementing the approach with subject-independent partitioning, we provided a more reliable assessment of generalization and highlighted how flawed protocols can also distort the evaluation of model compression techniques such as quantization. To ensure integrity and reproducibility, future research should adopt rigorous subject-aware evaluation protocols and openly share datasets and code.